Bias and Representativeness in Linked Data

各个领域的公司和组织以越来越快的速度收集和生成数据。描述单个实体（如人、产品）的数据可以由不同的组织或同一组织内的不同单位收集。数据不仅由不同的组织收集，而且可能以不同的方式和不同的粒度收集（例如，一家公司可能收集客户所在的城市，而另一家公司可能收集省份）。如今，人们普遍认为，通过整合来自不同来源的数据，我们可以丰富关于感兴趣实体的知识，从而使数据更有分析和预测的价值。虽然数据集成是一个具有挑战性的问题，但最近的技术进步使得创建大型数据湖成为可能，其中来自多个来源的数据是统一的。然而，我们离完全集成数据还很远（即，找到跨不同数据源的所有实体）。这是由于跨数据集合的唯一标识符不存在，因此必须对所有数据库使用共同特征，并比较它们的值以确定相似性。此外，不同的数据库模式、排版错误和丢失数据也会带来其他挑战。当从多个来源集成的数据用于分析或作为基于人工智能的方法的训练数据时，假阳性（不匹配）和假阴性（错过匹配）的影响是什么？拟议研究计划的主要目标是调查、理解和减轻通过数据集成产生的有偏见的数据。对数据偏差、代表性和质量的理解是至关重要的。当数据被用于可能影响整个社会的情况（例如，医疗保健、政策制定）时，尤其如此。为了实现所提出的计划的目标，研究计划分为一组短期目标和一组长期目标。短期目标如下：(1)调查目前用于数据集成和记录链接的最先进的系统；(2)建立一个比较框架，并提出新的方法来研究统一多个数据源生成的数据的代表性和偏差；(3)对不同类型实体（如人员、产品）的不同应用（如医疗保健、零售）中的技术和比较方法进行测试和评估。长期目标涉及开发新方法和新技术，以减轻偏见，并提高通过数据集成技术产生的数据的质量和价值。

Data are collected and generated at increasingly fast rates by companies and organizations in various domains. Data describing a single entity (e.g., person, product) can be collected by different organizations or by different units within the same organization. Not only that data are collected by different organizations, but it may be collected in different ways and at different granularity (e.g., one company may collect city for location of a customer, while another may collect province). Nowadays it is well accepted that by integrating data from different sources, we enrich the knowledge about entities of interest, thus making the data more valuable for analysis and prediction. Although data integration is a challenging problem, recent technological advances have made it possible to create large data lakes where data are unified from multiple sources. However, we are far away from integrating data fully (i.e., finding all the entities across the different data sources). This is due to the fact that unique identifiers across data collections do not exist, thus one must use common characteristics to all of the databases and compare their values to determine similarity. In addition, other challenges are presented by different database schemas, typographical errors and missing data. What is the impact of false positives (mismatches) and false negatives (missed matches) when data integrated from multiple sources are used for analysis or as training data for artificial intelligence -based methods? The main objective of the proposed research program is to investigate, understand and mitigate the biased data that are created through data integration. The understanding of the bias, representativeness and quality of data is critical. This is especially true when data are used in circumstances that could affect society at large (e.g., healthcare, policy making). Towards achieving the goal of the proposed program, the research plan is divided between a set of short- term goals and a set of long-term goals. The set of short -term goals are as follows: (1) investigating the state- of -the -art systems currently employed in data integration and record linkage; (2) developing a comparison framework and proposing new methods to investigate representativeness and bias in data generated by unifying multiple data sources; and (3) testing and evaluating the techniques and comparison methods in diverse applications (e.g., healthcare, retail) with different types of entities (e.g., persons, products). The long-term goal involves developing new methods and new technologies to mitigate the bias and to improve the quality and value of the data generated through data integration techniques.